fix(eval): emit string justification for coded evaluators

[ameyjain]

Problem

Coded evaluators (tool-call count/args/order/output, json-similarity) returned a null justification in their score output, while the LLM-judge evaluators returned a populated one.

Root cause: each coded evaluator stored its explanation under a per-evaluator key (explained_tool_calls_count, explained_tool_calls_args, explained_tool_calls_outputs, lcs, matched_leaves/total_leaves) and never under a justification key. The downstream eval worker reads details["justification"], which therefore always resolved to null — even though the structured detail was present.

Fix

Add a computed justification: str field to each of the five coded justification models, derived from the model's existing structured detail via a shared format_explained_tool_calls helper. model_dump() now emits a string justification for every coded evaluator, matching LLMJudgeJustification, without changing any of the structured fields.

Evaluator	justification now derived from
tool-call-count	explained_tool_calls_count
tool-call-args	explained_tool_calls_args
tool-call-output	explained_tool_calls_outputs
tool-call-order	lcs
json-similarity	matched_leaves / total_leaves

The base BaseEvaluatorJustification is intentionally left untouched: adding a computed justification there collides with LLMJudgeJustification's real justification field (pydantic raises TypeError). So the computed field lives on each coded subclass instead.

Verification

• pytest tests/evaluators tests/cli/eval — all pass
• mypy — clean
• ruff check + ruff format — clean

Related

Pairs with a python-eval-worker change that stops flattening these structured justifications, so the full object (including this justification string) reaches the client as justificationObject for per-evaluator-type rendering.

[sonarqubecloud]

Quality Gate failed

Failed conditions
67.7% Coverage on New Code (required ≥ 90%)

See analysis details on SonarQube Cloud